[ < ] [ > ]   [ << ] [ Up ] [ >> ]         [Top] [Contents] [Index] [ ? ]

D. Inline Assembler

If you need to write low-level software that interacts directly with the hardware, Ada provides two ways to incorporate assembly language code into your program. First, you can import and invoke external routines written in assembly language, an Ada feature fully supported by GNAT. However, for small sections of code it may be simpler or more efficient to include assembly language statements directly in your Ada source program, using the facilities of the implementation-defined package System.Machine_Code, which incorporates the gcc Inline Assembler. The Inline Assembler approach offers a number of advantages, including the following:

This chapter presents a series of examples to show you how to use the Inline Assembler. Although it focuses on the Intel x86, the general approach applies also to other processors. It is assumed that you are familiar with Ada and with assembly language programming.

D.1 Basic Assembler Syntax  
D.2 A Simple Example of Inline Assembler  
D.3 Output Variables in Inline Assembler  
D.4 Input Variables in Inline Assembler  
D.5 Inlining Inline Assembler Code  
D.6 Other Asm Functionality  


[ < ] [ > ]   [ << ] [ Up ] [ >> ]         [Top] [Contents] [Index] [ ? ]

D.1 Basic Assembler Syntax

The assembler used by GNAT and gcc is based not on the Intel assembly language, but rather on a language that descends from the AT&T Unix assembler as (and which is often referred to as "AT&T syntax"). The following table summarizes the main features of as syntax and points out the differences from the Intel conventions. See the gcc as and gas (an as macro pre-processor) documentation for further information.

Register names
gcc / as: Prefix with "%"; for example %eax
Intel: No extra punctuation; for example eax

Immediate operand
gcc / as: Prefix with "$"; for example $4
Intel: No extra punctuation; for example 4

Address
gcc / as: Prefix with "$"; for example $loc
Intel: No extra punctuation; for example loc

Memory contents
gcc / as: No extra punctuation; for example loc
Intel: Square brackets; for example [loc]

Register contents
gcc / as: Parentheses; for example (%eax)
Intel: Square brackets; for example [eax]

Hexadecimal numbers
gcc / as: Leading "0x" (C language syntax); for example 0xA0
Intel: Trailing "h"; for example A0h

Operand size
gcc / as: Explicit in op code; for example movw to move a 16-bit word
Intel: Implicit, deduced by assembler; for example mov

Instruction repetition
gcc / as: Split into two lines; for example
rep
stosl
Intel: Keep on one line; for example rep stosl

Order of operands
gcc / as: Source first; for example movw $4, %eax
Intel: Destination first; for example mov eax, 4


[ < ] [ > ]   [ << ] [ Up ] [ >> ]         [Top] [Contents] [Index] [ ? ]

D.2 A Simple Example of Inline Assembler

The following example will generate a single assembly language statement, nop, which does nothing. Despite its lack of run-time effect, the example will be useful in illustrating the basics of the Inline Assembler facility.

 
with System.Machine_Code; use System.Machine_Code;
procedure Nothing is
begin
   Asm ("nop");
end Nothing;

Asm is a procedure declared in package System.Machine_Code; here it takes one parameter, a template string that must be a static expression and that will form the generated instruction. Asm may be regarded as a compile-time procedure that parses the template string and additional parameters (none here), from which it generates a sequence of assembly language instructions.

The examples in this chapter will illustrate several of the forms for invoking Asm; a complete specification of the syntax is found in the GNAT Reference Manual.

Under the standard GNAT conventions, the Nothing procedure should be in a file named `nothing.adb'. You can build the executable in the usual way:
 
gnatmake nothing
However, the interesting aspect of this example is not its run-time behavior but rather the generated assembly code. To see this output, invoke the compiler as follows:
 
   gcc -c -S -fomit-frame-pointer -gnatp `nothing.adb'
where the options are:

-c
compile only (no bind or link)
-S
generate assembler listing
-fomit-frame-pointer
do not set up separate stack frames
-gnatp
do not add runtime checks

This gives a human-readable assembler version of the code. The resulting file will have the same name as the Ada source file, but with a .s extension. In our example, the file `nothing.s' has the following contents:

 
.file "nothing.adb"
gcc2_compiled.:
___gnu_compiled_ada:
.text
   .align 4
.globl __ada_nothing
__ada_nothing:
#APP
   nop
#NO_APP
   jmp L1
   .align 2,0x90
L1:
   ret

The assembly code you included is clearly indicated by the compiler, between the #APP and #NO_APP delimiters. The character before the 'APP' and 'NOAPP' can differ on different targets. For example, GNU/Linux uses '#APP' while on NT you will see '/APP'.

If you make a mistake in your assembler code (such as using the wrong size modifier, or using a wrong operand for the instruction) GNAT will report this error in a temporary file, which will be deleted when the compilation is finished. Generating an assembler file will help in such cases, since you can assemble this file separately using the as assembler that comes with gcc.

Assembling the file using the command

 
as `nothing.s'
will give you error messages whose lines correspond to the assembler input file, so you can easily find and correct any mistakes you made. If there are no errors, as will generate an object file `nothing.out'.


[ < ] [ > ]   [ << ] [ Up ] [ >> ]         [Top] [Contents] [Index] [ ? ]

D.3 Output Variables in Inline Assembler

The examples in this section, showing how to access the processor flags, illustrate how to specify the destination operands for assembly language statements.

 
with Interfaces; use Interfaces;
with Ada.Text_IO; use Ada.Text_IO;
with System.Machine_Code; use System.Machine_Code;
procedure Get_Flags is
   Flags : Unsigned_32;
   use ASCII;
begin
   Asm ("pushfl"          & LF & HT & -- push flags on stack
        "popl %%eax"      & LF & HT & -- load eax with flags
        "movl %%eax, %0",             -- store flags in variable
        Outputs => Unsigned_32'Asm_Output ("=g", Flags));
   Put_Line ("Flags register:" & Flags'Img);
end Get_Flags;

In order to have a nicely aligned assembly listing, we have separated multiple assembler statements in the Asm template string with linefeed (ASCII.LF) and horizontal tab (ASCII.HT) characters. The resulting section of the assembly output file is:

 
#APP
   pushfl
   popl %eax
   movl %eax, -40(%ebp)
#NO_APP

It would have been legal to write the Asm invocation as:

 
Asm ("pushfl popl %%eax movl %%eax, %0")

but in the generated assembler file, this would come out as:

 
#APP
   pushfl popl %eax movl %eax, -40(%ebp)
#NO_APP

which is not so convenient for the human reader.

We use Ada comments at the end of each line to explain what the assembler instructions actually do. This is a useful convention.

When writing Inline Assembler instructions, you need to precede each register and variable name with a percent sign. Since the assembler already requires a percent sign at the beginning of a register name, you need two consecutive percent signs for such names in the Asm template string, thus %%eax. In the generated assembly code, one of the percent signs will be stripped off.

Names such as %0, %1, %2, etc., denote input or output variables: operands you later define using Input or Output parameters to Asm. An output variable is illustrated in the third statement in the Asm template string:
 
movl %%eax, %0
The intent is to store the contents of the eax register in a variable that can be accessed in Ada. Simply writing movl %%eax, Flags would not necessarily work, since the compiler might optimize by using a register to hold Flags, and the expansion of the movl instruction would not be aware of this optimization. The solution is not to store the result directly but rather to advise the compiler to choose the correct operand form; that is the purpose of the %0 output variable.

Information about the output variable is supplied in the Outputs parameter to Asm:
 
Outputs => Unsigned_32'Asm_Output ("=g", Flags));

The output is defined by the Asm_Output attribute of the target type; the general format is
 
Type'Asm_Output (constraint_string, variable_name)

The constraint string directs the compiler how to store/access the associated variable. In the example
 
Unsigned_32'Asm_Output ("=m", Flags);
the "m" (memory) constraint tells the compiler that the variable Flags should be stored in a memory variable, thus preventing the optimizer from keeping it in a register. In contrast,
 
Unsigned_32'Asm_Output ("=r", Flags);
uses the "r" (register) constraint, telling the compiler to store the variable in a register.

If the constraint is preceded by the equal character (=), it tells the compiler that the variable will be used to store data into it.

In the Get_Flags example, we used the "g" (global) constraint, allowing the optimizer to choose whatever it deems best.

There are a fairly large number of constraints, but the ones that are most useful (for the Intel x86 processor) are the following:

=
output constraint
g
global (i.e. can be stored anywhere)
m
in memory
I
a constant
a
use eax
b
use ebx
c
use ecx
d
use edx
S
use esi
D
use edi
r
use one of eax, ebx, ecx or edx
q
use one of eax, ebx, ecx, edx, esi or edi

The full set of constraints is described in the gcc and as documentation; note that it is possible to combine certain constraints in one constraint string.

You specify the association of an output variable with an assembler operand through the %n notation, where n is a non-negative integer. Thus in
 
Asm ("pushfl"          & LF & HT & -- push flags on stack
     "popl %%eax"      & LF & HT & -- load eax with flags
     "movl %%eax, %0",             -- store flags in variable
     Outputs => Unsigned_32'Asm_Output ("=g", Flags));
%0 will be replaced in the expanded code by the appropriate operand, whatever the compiler decided for the Flags variable.

In general, you may have any number of output variables:

For example:
 
Asm ("movl %%eax, %0" & LF & HT &
     "movl %%ebx, %1" & LF & HT &
     "movl %%ecx, %2",
     Outputs => (Unsigned_32'Asm_Output ("=g", Var_A),   --  %0 = Var_A
                 Unsigned_32'Asm_Output ("=g", Var_B),   --  %1 = Var_B
                 Unsigned_32'Asm_Output ("=g", Var_C))); --  %2 = Var_C
where Var_A, Var_B, and Var_C are variables in the Ada program.

As a variation on the Get_Flags example, we can use the constraints string to direct the compiler to store the eax register into the Flags variable, instead of including the store instruction explicitly in the Asm template string:

 
with Interfaces; use Interfaces;
with Ada.Text_IO; use Ada.Text_IO;
with System.Machine_Code; use System.Machine_Code;
procedure Get_Flags_2 is
   Flags : Unsigned_32;
   use ASCII;
begin
   Asm ("pushfl"      & LF & HT & -- push flags on stack
        "popl %%eax",             -- save flags in eax
        Outputs => Unsigned_32'Asm_Output ("=a", Flags));
   Put_Line ("Flags register:" & Flags'Img);
end Get_Flags_2;

The "a" constraint tells the compiler that the Flags variable will come from the eax register. Here is the resulting code:

 
#APP
   pushfl
   popl %eax
#NO_APP
   movl %eax,-40(%ebp)

The compiler generated the store of eax into Flags after expanding the assembler code.

Actually, there was no need to pop the flags into the eax register; more simply, we could just pop the flags directly into the program variable:

 
with Interfaces; use Interfaces;
with Ada.Text_IO; use Ada.Text_IO;
with System.Machine_Code; use System.Machine_Code;
procedure Get_Flags_3 is
   Flags : Unsigned_32;
   use ASCII;
begin
   Asm ("pushfl"  & LF & HT & -- push flags on stack
        "pop %0",             -- save flags in Flags
        Outputs => Unsigned_32'Asm_Output ("=g", Flags));
   Put_Line ("Flags register:" & Flags'Img);
end Get_Flags_3;


[ < ] [ > ]   [ << ] [ Up ] [ >> ]         [Top] [Contents] [Index] [ ? ]

D.4 Input Variables in Inline Assembler

The example in this section illustrates how to specify the source operands for assembly language statements. The program simply increments its input value by 1:

 
with Interfaces; use Interfaces;
with Ada.Text_IO; use Ada.Text_IO;
with System.Machine_Code; use System.Machine_Code;
procedure Increment is

   function Incr (Value : Unsigned_32) return Unsigned_32 is
      Result : Unsigned_32;
   begin
      Asm ("incl %0",
           Inputs  => Unsigned_32'Asm_Input ("a", Value),
           Outputs => Unsigned_32'Asm_Output ("=a", Result));
      return Result;
   end Incr;

   Value : Unsigned_32;

begin
   Value := 5;
   Put_Line ("Value before is" & Value'Img);
   Value := Incr (Value);
   Put_Line ("Value after is" & Value'Img);
end Increment;

The Outputs parameter to Asm specifies that the result will be in the eax register and that it is to be stored in the Result variable.

The Inputs parameter looks much like the Outputs parameter, but with an Asm_Input attribute. The "=" constraint, indicating an output value, is not present.

You can have multiple input variables, in the same way that you can have more than one output variable.

The parameter count (%0, %1) etc, now starts at the first input statement, and continues with the output statements. When both parameters use the same variable, the compiler will treat them as the same %n operand, which is the case here.

Just as the Outputs parameter causes the register to be stored into the target variable after execution of the assembler statements, so does the Inputs parameter cause its variable to be loaded into the register before execution of the assembler statements.

Thus the effect of the Asm invocation is:

  1. load the 32-bit value of Value into eax
  2. execute the incl %eax instruction
  3. store the contents of eax into the Result variable

The resulting assembler file (with `-O2' optimization) contains:
 
_increment__incr.1:
   subl $4,%esp
   movl 8(%esp),%eax
#APP
   incl %eax
#NO_APP
   movl %eax,%edx
   movl %ecx,(%esp)
   addl $4,%esp
   ret


[ < ] [ > ]   [ << ] [ Up ] [ >> ]         [Top] [Contents] [Index] [ ? ]

D.5 Inlining Inline Assembler Code

For a short subprogram such as the Incr function in the previous section, the overhead of the call and return (creating / deleting the stack frame) can be significant, compared to the amount of code in the subprogram body. A solution is to apply Ada's Inline pragma to the subprogram, which directs the compiler to expand invocations of the subprogram at the point(s) of call, instead of setting up a stack frame for out-of-line calls. Here is the resulting program:

 
with Interfaces; use Interfaces;
with Ada.Text_IO; use Ada.Text_IO;
with System.Machine_Code; use System.Machine_Code;
procedure Increment_2 is

   function Incr (Value : Unsigned_32) return Unsigned_32 is
      Result : Unsigned_32;
   begin
      Asm ("incl %0",
           Inputs  => Unsigned_32'Asm_Input ("a", Value),
           Outputs => Unsigned_32'Asm_Output ("=a", Result));
      return Result;
   end Incr;
   pragma Inline (Increment);

   Value : Unsigned_32;

begin
   Value := 5;
   Put_Line ("Value before is" & Value'Img);
   Value := Increment (Value);
   Put_Line ("Value after is" & Value'Img);
end Increment_2;

Compile the program with both optimization (`-O2') and inlining enabled (`-gnatpn' instead of `-gnatp').

The Incr function is still compiled as usual, but at the point in Increment where our function used to be called:

 
pushl %edi
call _increment__incr.1

the code for the function body directly appears:

 
movl %esi,%eax
#APP
   incl %eax
#NO_APP
   movl %eax,%edx

thus saving the overhead of stack frame setup and an out-of-line call.


[ < ] [ > ]   [ << ] [ Up ] [ >> ]         [Top] [Contents] [Index] [ ? ]

D.6 Other Asm Functionality

This section describes two important parameters to the Asm procedure: Clobber, which identifies register usage; and Volatile, which inhibits unwanted optimizations.

D.6.1 The Clobber Parameter  
D.6.2 The Volatile Parameter  


[ < ] [ > ]   [ << ] [ Up ] [ >> ]         [Top] [Contents] [Index] [ ? ]

D.6.1 The Clobber Parameter

One of the dangers of intermixing assembly language and a compiled language such as Ada is that the compiler needs to be aware of which registers are being used by the assembly code. In some cases, such as the earlier examples, the constraint string is sufficient to indicate register usage (e.g., "a" for the eax register). But more generally, the compiler needs an explicit identification of the registers that are used by the Inline Assembly statements.

Using a register that the compiler doesn't know about could be a side effect of an instruction (like mull storing its result in both eax and edx). It can also arise from explicit register usage in your assembly code; for example:
 
Asm ("movl %0, %%ebx" & LF & HT &
     "movl %%ebx, %1",
     Inputs  => Unsigned_32'Asm_Input  ("g", Var_In),
     Outputs => Unsigned_32'Asm_Output ("=g", Var_Out));
where the compiler (since it does not analyze the Asm template string) does not know you are using the ebx register.

In such cases you need to supply the Clobber parameter to Asm, to identify the registers that will be used by your assembly code:

 
Asm ("movl %0, %%ebx" & LF & HT &
     "movl %%ebx, %1",
     Inputs  => Unsigned_32'Asm_Input  ("g", Var_In),
     Outputs => Unsigned_32'Asm_Output ("=g", Var_Out),
     Clobber => "ebx");

The Clobber parameter is a static string expression specifying the register(s) you are using. Note that register names are not prefixed by a percent sign. Also, if more than one register is used then their names are separated by commas; e.g., "eax, ebx"

The Clobber parameter has several additional uses:

  1. Use "register" name cc to indicate that flags might have changed
  2. Use "register" name memory if you changed a memory location


[ < ] [ > ]   [ << ] [ Up ] [ >> ]         [Top] [Contents] [Index] [ ? ]

D.6.2 The Volatile Parameter

Compiler optimizations in the presence of Inline Assembler may sometimes have unwanted effects. For example, when an Asm invocation with an input variable is inside a loop, the compiler might move the loading of the input variable outside the loop, regarding it as a one-time initialization.

If this effect is not desired, you can disable such optimizations by setting the Volatile parameter to True; for example:

 
Asm ("movl %0, %%ebx" & LF & HT &
     "movl %%ebx, %1",
     Inputs   => Unsigned_32'Asm_Input  ("g", Var_In),
     Outputs  => Unsigned_32'Asm_Output ("=g", Var_Out),
     Clobber  => "ebx",
     Volatile => True);

By default, Volatile is set to False unless there is no Outputs parameter.

Although setting Volatile to True prevents unwanted optimizations, it will also disable other optimizations that might be important for efficiency. In general, you should set Volatile to True only if the compiler's optimizations have created problems.


[ << ] [ >> ]           [Top] [Contents] [Index] [ ? ]

This document was generated by Mail Server on June, 15 2005 using texi2html